news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

the latest ai vincentian graph modelFlux1.1, refreshed the screen overnight.

with just one simple trick, you canremove the "ai smell" from the picture, both people and landscapes can achieve photo-quality effects.

the reaction of netizens in the comment area was like: i can’t tell the difference, i really can’t tell the difference.

this technique is also very simple to use. in the prompt wordimitate the file naming format of an slr camerathat’s it.

for example“CR2”it is the raw image file format used by canon cameras, input"img" + random number + ".cr2", plus the specified content, you can get a realistic image.

later, there were also feedbacks from netizens who tried it and switched to sony cameras.“ARW”, nikon camera“NEF”, even apple’s“HEIC”format, you can get good results.

so much so that some people began to doubt that the model randomly spit out a real photo from the training data, right?

however, when you zoom in and look at some specific details, it is easy to see that it is indeed generated by ai. for example, the text on the license plate number is garbled.

so is the flux1.1 model itself very powerful? how much role does this technique play in it?

a senior photo retouching artist posted a comparison. he added img_1018.cr2 on the left and without it on the right. he thought the difference was huge.

it can also be seen from our actual measurement results that adding this technique can significantly improve the authenticity of the picture.

now if you want to try this flux1.1 model for free, you can come totogether.aiplatform, you will get 5 usd points when you sign up.

randomly generate a selfie of a tourist at the great wall. at first glance, it looks like that, but if you look closely at the character's skin texture, background mountains and plants, it still smells like ai.

if you replace it with "img_0314.cr2: selfie on the great wall", will it be different immediately?

codenamed blueberry, the latest sota vincent diagram model

with the official release of flux1.1, two unclaimed models that have repeatedly topped the list of vincentian models have"blueberry"the mystery of the model is also lifted, it is exactly that.

the official no longer hides it and releases the data directly on the artificial analysis image arena.FLUX1.1 [pro]codenamed "blueberry," it surpassed all other models and achieved the highest overall elo score.

in comparison, flux1.1 [pro]it’s also cheaper and faster, various indicators surpass midjourney, sd3, ideogram, etc.

in terms of generation speed, flux1.1 [pro] is 6 times faster than the previous generation flux.1 [pro] while maintaining image quality, command response and diversity improvements.

by the way, now flux.1 [pro] has also been updated and is 2 times faster than before, flux1.1 [pro] is 3 times faster than the currently available flux.1 [pro].

in addition, the official said that fast high-resolution generation will be launched soon, the kind that flux1.1 [pro] can natively support, and can generate 2k images without sacrificing any command response.

flux1.1 [pro] will be available through online platforms such as together.ai, replicate, fal.ai, and freepik.

at the same time, the official also launched the bfl api, which can be integrated by other developers into their own applications. the api pricing is:

flux.1 [dev]: 2.5 cents per picture (approximately rmb 0.18 yuan)

flux.1 [pro]: 5 cents per picture (approximately rmb 0.35 yuan)

flux1.1 [pro]: 4 cents per picture (approximately rmb 0.28 yuan)

created by the original team of stable diffusion

behind flux1.1 [pro] isstable diffusion original team, members include robin rombach, andreas blattmann, dominik lorenz, etc.

△Robin Rombach

in fact, stable diffusion was originally an academic research project.

led by professor björn ommer, several members of the machine vision and learning research group at the university of munich, including robin rombach, andreas blattmann, dominik lorenz, and runway researcher patrick esser.

seven months after the research paper was published, stability ai stepped in to provide computing resources to further develop the text-to-image generation model. in 2022, the authors of the above papers will join stability ai.

together, the team created stable diffusion xl, stable video diffusion, etc.

one of the best papers of icml 2024, the rectified flow transformers of the stable diffusion 3 technical paper, and the adversarial diffusion distillation method used by sdxl-turbo are also being studied by this group of people.

in march this year, it was revealed that these core research team members resigned en masse.

then, they formed a new team calledBlack Forest Labs(black forest laboratories), headquartered in germany.

it was just announced in early august this year and released its first generation vincent graph model flux.1. flux.1 has three variants: flux.1 [pro], flux.1 [dev] and flux.1 [schnell], balanced performance and accessibility.

at present, black forest laboratory has completed a seed round of financing, with a total ofraised us$31 million, andreessen horowitz led the investment, with brendan iribe, michael ovitz, garry tan, timo aila and vladlen koltun among other investors.

it is said that they have also received follow-up investment from general catalyst and mätchvc.

black forest laboratorythere is still cooperation with musk, introducing its image generation model into xai’s grok assistant.

next, the team revealed that it will launch sota leveltext-to-video generative model。

it is said that theyis raising $100 million at a $1 billion valuation, a significant increase from the previous valuation of $150 million.

from pika 1.5 to meta movie gen, the video generation track exploded in the second half of this year, and the addition of black forest lab may bring a different spark.